text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
The problem with this approach is that if a large magnitude value (i.e., an outlier) occurs in the input
tensor, then the quantization bins—certain bit combinations—are not utilized well with few or no
numbers quantized in some bins. To prevent the outlier issue, a common approach is to chunk the
input tensor into bloc... | QLORA |
tests cases. An example correct solution generated by AlphaCode for the problem in Figure 2 is given
in Figure 3, and extensive results and analysis can be found in Section 5 and 6. | alphacode |
4.4 Results of Interactive Chat
We showcase the conversational capabilities of Qwen-Audio-Chat through illustrative cases depicted in
Figure 2. Furthermore, we intend to provide public access to the trained models for online chat interactions. | Qwen-Audio |
Human-like Open-Domain Chatbot. cs.CL.
Alcorn, M. A., Li, Q., Gong, Z., Wang, C., Mai, L., Ku, W.-S. et al. (2018). Strike (with) a Pose: Neural
Networks Are Easily Fooled by Strange Poses of Familiar Objects. arXiv, 1811.11553v3.
Arabshahi, F., Lu, Z., Singh, S., & Anandkumar, A. (2019). Memory Augmented Recursi... | The Next Decade in AI- |
burger meals: 4 * 11 = $44 6. Calculate the cost of 4 special burger meals: 4 * 9.50 = $38 7. Calculate the cost of 2 kid’s
burger meals: 2 * 7 = $14 8. Calculate the cost of 2 special kid’s burger meals: 2 * 5 = $10 9. Calculate the total savings:
savings on special burger meals + savings on kid’s burger meals = 6 + 4... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
7 Inductive logic programming (Cropper, Morel, & Muggleton, 2019) is a purely-rule based approach to
learning that is worth some consideration, though outside the scope of the current paper.
8 Although I am fairly confident that robust intelligence will depend on some sort of hybrid that
combines symbolic operations... | The Next Decade in AI- |
B. Storage Requirements
When applying our textual bypass, our mapper net-
works contain approximately 560, 000 learnable parame-
ters. When textual bypass is not applied, this reduces to
approximately 460, 000 trainable parameters. This amounts
to 2.2MB and 1.86MB of disk space required to store each
learned concept, ... | A Neural Space-Time Representation for Text-to-Image Personalization |
superior performance compared to the other two models. [3] showed that LLMs perform worse
on physics problems than chemistry problems, probably because chemistry problems have lower
inference complexity than physics problems in this setting. There are limited evaluation studies on
LLMs in the field of general science, ... | ASurveyonEvaluationofLargeLanguageModels |
We use the Stable Diffusion [44] as an example to introduce the method to use ControlNet to control
a large diffusion model with task-specific conditions.
Stable Diffusion is a large text-to-image diffusion model trained on billions of images. The model
is essentially an U-net with an encoder, a middle block, and a skip... | Adding Conditional Control to Text-to-Image Diffusion Models |
Planning with Large Language Models Various large language models (LLMs) have been devel-
oped in recent years, such as Bert [27], CodeX [28], Opt [29], GPT-3 [10], ChatGPT [30], GPT-4 [2],
LLAMA [31]. and PaLM [32]. As LLMs are pretrained with a tremendous amount of offline text
data, they can emerge with surprising ze... | LLM+P- Empowering Large Language Models with Optimal Planning Proficiency |
ety of criteria compared with existing music
generation models. Lastly, to promote the open-
source culture, we provide a collection of open-
source libraries with the hope of facilitating
future work in the field.1 | MOUSAI |
1Models large enough to achieve good factual coverage
require extreme amounts of compute, and the largest neural
LMs now cost millions of dollars to train (Brown et al., 2020). | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
3Supervised PWC is simply an ensemble version of the clas-
sic method (Gray and Moore, 2003; Ram and Gray, 2011; Wu
et al., 2014). To the best of our knowledge, no one has previously
proposed unsupervised PWC density estimation with CART trees.
This can be understood as a variant of our approach in which all
marginals ... | Adversarial Random Forests for Density Estimation and Generative Modeling |
A Review of Deep Learning Techniques for Speech Processing
109
(2015).
[620] Yusuke Yasuda, Xin Wang, Shinji Takaki, and Junichi Yamagishi. 2019. Investigation of Enhanced Tacotron Text-to-
speech Synthesis Systems with Self-attention for Pitch Accent Language. In ICASSP 2019 - 2019 IEEE International
Conference on ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
To overcome the locality of the analytical gradient of
hash encoding, we propose to compute the surface normals
using numerical gradients. If the step size of the numeri-
cal gradient is smaller than the grid size of hash encoding,
the numerical gradient would be equivalent to the analyti-
cal gradient; otherwise, hash... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
sha1_base64="76w10YEtETzUXdaT0wTZt0xBig8=">AAAB9XicbVDLSgMxFL1TX7W+qi7dBIvgqsyIoMuCG5cV7EPaacmkmTY0kxmSO0oZ+h9uXCji1n9x59+YtrPQ1gOBwzn3ck9OkEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmjjVjDdYLGPdDqjhUijeQIGStxPNaRRI3grGNzO/9ci1EbG6x0nC/YgOlQgFo2ilXjeiOArCrD3tYV/0yxW36s5BVomXkwrkqPfLX91BzNKIK2SSGtPx3AT9jGoUTPJpqZsanlA2pkPesVTRiBs/m... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
In addition to observing that participants can change
their mind after viewing the flagged tweets, we found indi-
vidual differences also influenced the likelihood partici-
pants would change their mind after exposure to the flags.
Individual attitudes such as anomie (the view that there is
a societal breakdown in n... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
We quantitatively compare our method with the state-
of-the-art methods using both Twindom testing dataset and
BUFF rendering dataset to evaluate the geometry recon-
sturction accuracy. Similar to the experiments in PIFu [8], we
use point-to-surface error as well as the Chamfer distance
as error metric. The numerical r... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
an end-to-end fashion, achieving superior performance in knowledge-intensive NLP tasks (Guu et al., 2020;
Lewis et al., 2020b; Izacard et al., 2022).
Later works have gone beyond local repositories, for instance, some leverage the entire web as the knowledge
source, which allows for improved temporal generalization and... | Tool Learning with Foundation Models |
language understanding. Advances in neural information processing systems, 32, 2019.
[120] Wenpeng Yin, Jamaal Hay, and Dan Roth. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. In Proceedings
of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
including an objective experiment, as well as a subjective listening study, which
shows that our proposed Video2Music framework is able to successfully generate
music that matches video, with a quality that outperforms the state-of-the-art.
In sum, our music generation system represents a pioneering approach to
tac... | Video2Music |
Sequence
Per CS-2
Length Batch Size
121
33
121
85
50
65
50
2,048
10,000
2,048
2,048
2,048
2,048
2,048
Performance relative to 1 CS-2
2 CS-2s
1.99x
1.99x
1.98x
1.99x
1.92x
1.97x
1.98x
4 CS-2s
3.94x
3.97x
3.91x
3.89x
3.75x
3.65x
3.92x
8 CS-2s
7.87x
7.95x
7.86x
7.91x
7.93x
7.69x
8.05x
16 CS-2s
15.50x
15.87x
15.62x
... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work ... | Generative Agents- Interactive Simulacra of Human Behavior |
5 Adding guardrails for front-facing applications
The ability to enforce guardrails when it comes to AI generation is important for front-facing appli-
cations. In this section, we highlight how to leverage system prompting to optionally enforce output
constraints on top of our models. Additionally, we showcase the ab... | Mistral7B |
preprint arXiv:2202.05008, 2022.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Infor-
mation Processing Systems, 2017. URL https://proceedings.neurips.cc/paper_
files/paper/2017/file/3f5ee24... | JAXPRUNER |
[52] Omar Khattab, Keshav Santhanam, Xiang Lisa Li, David Hall, Percy Liang,
Christopher Potts, and Matei Zaharia. 2023. Demonstrate-Search-Predict:
Composing retrieval and language models for knowledge-intensive NLP.
arXiv:2212.14024 [cs.CL]
[53] Bjoern Knafla. 2011. Introduction to Behavior Trees. http://bjoernknafl... | Generative Agents- Interactive Simulacra of Human Behavior |
3.1 BACKGROUND: PROBABILISTIC CIRCUITS
Probabilistic Circuits (PCs) are an umbrella term for a wide
variety of Tractable Probabilistic Models (TPMs). They pro-
vide a set of succinct definitions for popular TPMs such as
Sum-Product Networks (Poon & Domingos, 2011), Arithmetic
Circuits (Shen et al., 2016), and Probabili... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
49 Santa Clara University School of Law, 2018. Content Moderation & Removal at Scale con-
ference, Santa Clara, CA, February 2. https://law.scu.edu/event/content-moderation-removal-
at-scale. In one widely reported session, Emma Llansó of the Center for Democracy and
Technology and Mike Masnick of the blog Techdirt inv... | Social_Media_and_Democracy |
3.2.1 Human Preference Data Collection
Next, we collect human preference data for reward modeling. We chose a binary comparison protocol over
other schemes, mainly because it enables us to maximize the diversity of collected prompts. Still, other
strategies are worth considering, which we leave for future work.
Our ann... | Llama2 |
Choenni, R., Shutova, E., and van Rooij, R. Stepmothers are
mean and academics are pretentious: What do pretrained
language models learn about you? In Proceedings of
the 2021 Conference on Empirical Methods in Natural
Language Processing, pp. 1477–1491, 2021.
Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra,
G... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
B Implementation Details
B.1 Unsupervised pre-training
For unsupervised pre-training we build on the DINO and iBOT codebases. We use hyperparameters shown
in Table 16, ViT architectures described in Table 17.
KoLeo regularization. We apply the KoLeo regularizer with a weight of 0.1 between the class tokens of
the fir... | DINOv2- Learning Robust Visual Features without Supervision |
Diederik Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International
conference on learning representations (ICLR), San Diego, CA, USA, 2015. tex.optmonth: 12.
Nayoung Lee, Kartik Sreenivasan, Jason D. Lee, Kangwook Lee, and Dimitris Papailiopoulos.
Teaching Arithmetic to Small Transformers, Jul... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
E COMBINING LORA WITH PREFIX TUNING
LoRA can be naturally combined with existing prefix-based approaches. In this section, we evaluate
two combinations of LoRA and variants of prefix-tuning on WikiSQL and MNLI.
LoRA+PrefixEmbed (LoRA+PE) combines LoRA with prefix-embedding tuning, where we insert
lp + li special tokens wh... | LORA |
pattern for the AnEM dataset in Table 3 but only if we consider the global F1 score. However, when
engaging with a more fine-grained analysis we see that the error decreased on seen entities, but
increased on unseen entities. This highlights the need for more detailed evaluation processes to find
emergent differences bet... | MULTI HASH EMBEDDINGS IN SPACY |
Angela Fan Melanie Kambadur Sharan Narang Aurelien Rodriguez Robert Stojnic
Sergey Edunov Thomas Scialom∗
GenAI, Meta
Abstract
In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned
large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters.
Our fine-tuned... | Llama2 |
achieve sufficient accuracy on 1 through N digit addition before N + 1 digit examples are added to
the dataset, starting with N = 3. Accuracy is measured by computing an exact token match between
the gold reference text and the model output while sampling at temperature 0. Any answer that is
not in the correct format i... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
Deep Ganguli, Amanda Askell, Nicholas Schiefer, Thomas Liao, Kamil˙e Lukoši¯ut˙e, Anna Chen,
Anna Goldie, Azalia Mirhoseini, Catherine Olsson, Danny Hernandez, et al. The capacity for
moral self-correction in large language models. arXiv e-prints, pages arXiv–2302, 2023.
Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wa... | Self-AlignmentwithInstructionBacktranslation |
r
e
n
e
i
t
h
e
r
i
n
f
o
r
m
a
t
i
o
n
d
a
t
a
b
a
s
e
s
n
o
r
d
e
t
e
r
m
i
n
i
s
t
i
c
i
n
f
o
r
m
a
t
i
o
n
r
e
t
r
i
e
v
a
l
s
y
s
t
e
m
s
.
S
o
w
h
i
l
e
a
u
s
e
r
c
a
n
e
x
p
e
c
t
e
x
a
c
t
l
y
t
h
e
s
a
m
e
a
n
d
c
o
n
s
i
s
t
e
n
t
r
e
s
p
o
n
s
e
t
o
a
d
a
t
a
b
... | An overview of Bard- an early experiment with generative AI |
of data collection. Though we mostly collected labels on incorrect solutions, we
still collected many labels for correct individual steps. In fact, our small-scale
ablations in Section 4.2 suggest that this active learning strategy, which favors
labelling high-scoring wrong-answer solutions, improves performance despit... | Let’s Verify Step by Step |
prize: Second round winners, 2023.
arXiv preprint arXiv:1602.06023, 2016.
[72] Shashi Narayan, Shay B Cohen, and Mirella Lapata. Don’t give me the details, just the summary! topic-aware convolutional neural networks for
extreme summarization. arXiv preprint arXiv:1808.08745, 2018.
[73] Tri Nguyen, Mir Rosenberg, Xi... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
that any special requirements can be put in place.
Further guidance
1. Applicants with disabilities should contact the Disability, Mental Health and Wellbeing team in
Student Support and Wellbeing (SSW) if they have any general queries about facilities at UCL
before submitting their application.
2. UCL endeav... | UCL Academic Manual |
Annotation To validate this quantitatively, we con-
ducted a listener test with three perceivers (annota-
tors) with diverse demographic backgrounds (both
female and male, all with at least a Bachelor’s de-
gree of education). Each annotator listens to all
80 music samples we provide, and is instructed to
categorize ea... | MOUSAI |
3.4 Teacher-Student Architecture Specific Tricks
3.4.1 Role of the Moving Average Teacher
While the original BYOL method is based on exponential moving average (EMA) updates
of the weights for the target (teacher) network, it was later confirmed that EMA is not
necessary (i.e., the online and target networks can be ident... | A Cookbook of Self-Supervised Learning |
2018. URL https://arxiv.org/abs/1804.04235.
Renee Shelby, Shalaleh Rismani, Kathryn Henne, AJung Moon, Negar Rostamzadeh, Paul Nicholas, N’Mah
Yilla, Jess Gallegos, Andrew Smart, Emilio Garcia, and Gurleen Virk. Sociotechnical harms: Scoping a
taxonomy for harm reduction, 2022. URL https://arxiv.org/abs/2210.05791.
F... | Scaling Instruction-Finetuned Language Models |
3.3 Social Science
Social science involves the study of human society and individual behavior, including economics,
sociology, political science, law, and other disciplines. Evaluating the performance of LLMs in
social science is important for academic research, policy formulation, and social problem-solving.
Such eval... | ASurveyonEvaluationofLargeLanguageModels |
idiosyncrasies of the dataset, and their ability to generalize
robustly to out-of-distribution data could even degrade.
To check whether this is the case, we study the zero-shot
generalization of Whisper models as a function of the model
size. Our analysis is summarized in Figure 8. With the
exception of English speech... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
combination of the three elements: 𝑠𝑐𝑜𝑟𝑒 = 𝛼𝑟𝑒𝑐𝑒𝑛𝑐𝑦 · 𝑟𝑒𝑐𝑒𝑛𝑐𝑦 +
𝛼𝑖𝑚𝑝𝑜𝑟𝑡𝑎𝑛𝑐𝑒 · 𝑖𝑚𝑝𝑜𝑟𝑡𝑎𝑛𝑐𝑒 + 𝛼𝑟𝑒𝑙𝑒𝑣𝑎𝑛𝑐𝑒 · 𝑟𝑒𝑙𝑒𝑣𝑎𝑛𝑐𝑒. In our implemen-
tation, all 𝛼’s are set to 1. The top-ranked memories that fit in the
language model’s context window are then included in the ... | Generative Agents- Interactive Simulacra of Human Behavior |
applications begin to enter the fray.
Wave 3: Better, faster, cheaper (2022+) Compute gets cheaper. New techniques, like diffusion models, shrink down the costs required to train and run inference. The research community continues to develop better algorithms and larger models. Developer access expands from closed be... | Generative AI A Creative New World Sequoia Capital |
We find that training our models on the true variational bound yields better codelengths than training
on the simplified objective, as expected, but the latter yields the best sample quality. See Fig. 1 for
CIFAR10 and CelebA-HQ 256 × 256 samples, Fig. 3 and Fig. 4 for LSUN 256 × 256 samples [71],
and Appendix D for more... | Denoising Diffusion Probabilistic Models |
evaluations for those models. We notice that when downmixing the stereo output to mono, we are
almost equivalent in perceived quality to a mono model. Stereo audio was overall rated higher than
the mono counterpart, and the “stereo partial delay” benefits from a small boost both in overall quality
and text relevance co... | Simple and Controllable Music Generation |
12
Table 12: Trade-off between latency and WER performance with decreasing model size. Aver-
age WER over the 11 ID and three OOD validation sets as the number of encoder and decoder layers
in the large-v2 checkpoint are reduced. The first row corresponds to the teacher checkpoint large-v2.
The following rows corresp... | DISTIL-WHISPER |
12
Competition-Level Code Generation with AlphaCode
4.5. Filtering
To accurately represent competitive programming contests and penalties, our formulation limits us
to just 10 submissions per problem no matter how many samples we draw. One powerful tool for
selecting these submissions is filtering samples to only tho... | alphacode |
is admissible for c if 0 ≤ h(s, t) ≤ c(s, t), for all s, t ∈ S and h is consistent for c if h(s, t) ≤ c(s, u) + h(u, t) for all s, t, u ∈ S. It
is also common to instead define the heuristic function with respect to a particular set G of goal states as a function hG of
one state, i.e. hG (s) = mint∈G h(s, t). Our var... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
computing frameworks, such as edge computing and federated learning, in which cloud servers are responsible
for hosting computationally intensive models, while edge devices like PCs or smartphones process personalized
information to prevent its leakage. | Tool Learning with Foundation Models |
2022-3-16
Competition-Level Code Generation with
AlphaCode
Yujia Li*, David Choi*, Junyoung Chung*, Nate Kushman*, Julian Schrittwieser*, Rémi Leblond*, Tom
Eccles*, James Keeling*, Felix Gimeno*, Agustin Dal Lago*, Thomas Hubert*, Peter Choy*, Cyprien de
Masson d’Autume*, Igor Babuschkin, Xinyun Chen, Po-Sen Huang, ... | alphacode |
system having been tested extensively before launch (Roose,
2023; Perrigo, 2023; Mehdi, 2023).
Though there has been some progress in understanding and
mitigating these issues, there is no consensus on whether
or how we will be able to deeply solve them, and there is
increasing concern that they will become catastrophi... | Eight Things to Know about Large Language Models |
of the problem, simplify-then-guess asks the model to directly guess the final solution without using
any further reasoning steps. The final answer is a majority vote over all intermediate guesses. For
example, if a model is tasked with solving an 8-digit addition problem, it will first simplify the
problem into a 7 di... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
[631] Tsipras, D., S. Santurkar, L. Engstrom, et al. Robustness may be at odds with accuracy. In 7th
International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA,
May 6-9, 2019. OpenReview.net, 2019.
[632] Zhang, H., Y. Yu, J. Jiao, et al. Theoretically principled trade-off between robustness ... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
text as “multimodal sentences” of latent vectors, allowing
it to process multiple images in a flexible way within any
part of a sentence. More closely related to our work is
Frozen (Tsimpoukelli et al., 2021) where vision encoder
parameters are optimized via backpropagation through a
frozen LLM (Lu et al., 2021). Inspir... | PaLM-E- An Embodied Multimodal Language Model |
I take as my paradigm a certain type of human cognition—the type, for example, involved in e.g.,
planning and then taking a trip from New York to San Francisco; reasoning about the safest way
to cut down a tree, then doing it; designing a component of a particle collider; and so on. When I
talk about agentic planning, ... | Is Power-Seeking AI an Existential Risk? |
We used a single-channel image for the thermal data
since it is the natural form in which current infrared thermal
sensors return data [31]. For single-view depth, we ex-
perimented with different encodings – absolute depth [64]
as returned by sensors like the Kinect, inverse depth [61],
disparity [61], and HHA [24, 25... | IMAGEBIND- One Embedding Space To Bind Them A |
φ(n) of PC units n, that is, the collection of variables defined by all its descendent input units.
Definition 2 (Decomposability). A PC is decomposable if for every product unit n, its children have
disjoint scopes: ∀c1, c2 ∈ in(n) (c1 (cid:54)= c2), φ(c1) ∩ φ(c2) = ∅.
All product units in Fig. 1 are decomposable. For e... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
7.1 Encoder Experts Exhibit Specialization . . . . . . . . . . . . . . . . . . . . . . .
7.2 Decoder Experts Lack Specialization . . . . . . . . . . . . . . . . . . . . . . . . .
7.3 Multilingual Experts Specialize, But Not by Language
. . . . . . . . . . . . . . .
8 Related Work
9 Discussion
10 Conclusion
A Token Loa... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
3. Completed a minimum of eighteen months of work experience no more than two years prior to
the proposed date of enrolment, evidenced by a letter from the employer including start and end
dates and language of business, in one of the following countries:
Antigua and Barbuda, Australia, Barbados, Belize, Botswana... | UCL Academic Manual |
We're also working to add capabilities to Med-PaLM 2, so that it can synthesize information from medical imaging
like plain films and mammograms. You can imagine an AI collaborator that helps radiologists interpret images and
communicate the results. These are some examples of PaLM 2 being used in specialized domains. ... | Google I_O 2023_ Making AI more helpful for everyone |
include a copy of the advertisement, targeting data, as well as information
about the purchaser of the advertisement.43 Similar approaches outside the
advertising context might attempt
to prevent bots and astroturfing by
mandating more stringent requirements on the creation of new user accounts
and profiles on a service. | Social_Media_and_Democracy |
Not all forms of participation, of course, are equally benign or pro-social. The
Internet, and the use we make of it, is profoundly ambiguous. Various forms of
online harassment and trolling once thought to be relatively marginal and
subcultural phenomena are now mainstream and widely experienced, enabled
by digital te... | Social_Media_and_Democracy |
e.g., by checking whether it hallucinates evidence.
If so, LLM-Augmenter generates a feedback mes-
sage. The message is used to revise the prompt to
query GPT-3.5 again. The process iterates until a
candidate response passes the verification and is
sent to the user.
FreshPrompt: (Vu et al., 2023) address the static
nat... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
magnitude weights). Unstructured pruning produces a model with ”holes” or sparse
weight matrices, which require specialized software or hardware for efficient deploy-
ment. Recent research efforts have been devoted to combining LLMs with pruning
techniques, aiming to tackle the substantial size and computational costs ass... | Beyond Efficiency |
In conclusion, we are committed to conducting
our research responsibly and ethically. We encour-
age the research community to engage in open dis-
cussions about the ethical implications of text-to-
music generation models and to develop guidelines
and best practices for their responsible use. By
addressing these conce... | Moûsai |
detection and indicative of outside efforts to influence digital political
conversation in other countries (Monaco 2017; Varol et al. 2017). Owing to
the complexity of how networks of political bots operate – with the same
collections of accounts switching focus between state borders and across
multiple tongues – they a... | Social_Media_and_Democracy |
is targeted to learn a mapping between images in different domains, while a ControlNet is targeted to
control a diffusion model with task-specific conditions.
Pix2Pix [20] presented the concept of image-to-image translation, and early methods are dominated
by conditional generative neural networks [20, 69, 60, 39, 8, 63... | Adding Conditional Control to Text-to-Image Diffusion Models |
Introduction
1
In recent years, natural language processing (NLP)
has made significant strides in understanding and
generating human language, due to the advance-
ments in deep learning and large-scale pre-trained
models (Radford et al., 2018; Devlin et al., 2019;
Brown et al., 2020). While the majority of NLP
researc... | Moûsai |
sthepolicyforTtimesteps(whereTismuchlessthantheepisodelength),andusesthecollectedsamplesforanupdate.ThisstylerequiresanadvantageestimatorthatdoesnotlookbeyondtimestepT.Theestimatorusedby[Mni+16]isˆAt=−V(st)+rt+γrt+1+···+γT−t+1rT−1+γT−tV(sT)(10)wheretspecifiesthetimeindexin[0,T],withinagivenlength-Ttrajectorysegment.Gene... | PPO |
Fine-tuning setting
No Enhancements
19.6% (18.2-20.4)
+ MLM
20.7% (19.1-21.3)
+ Tempering
21.9% (20.7-22.6)
+ Tags and Ratings
22.4% (21.3-23.0)
+ Value
23.2% (21.7-23.9)
+ GOLD
24.2% (23.1-24.4)
+ Clustering
28.4% (27.5-29.3)
Table 8 | Build-up ablation for model enhancements. Effect of each additional model enhancemen... | alphacode |
2
2
0
2
c
e
D
7
]
L
C
.
s
c
[
1
v
3
3
5
3
0
.
2
1
2
2
:
v
i
X
r
a
Text Embeddings by Weakly-Supervised
Contrastive Pre-training
Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao
Linjun Yang, Daxin Jiang, Rangan Majumder, Furu Wei
Microsoft Corporation
https://github.com/microsoft/unilm
Abstract | E5 |
Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014.
[57] Stuart J Russell. Artificial intelligence a modern approach. Pearson Education, Inc., 2010.
[58] William Saunders, Catherine Yeh, Jeff Wu, Steven Bills, Long Ouyang, Jonathan Ward, and Jan Leike.
Self-critiquing models for assisting ... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
2 RELATED WORK | DISTIL-WHISPER |
3. The applicant will be informed in writing by the Director of Access and Admissions of the
apparent misrepresentation and asked to provide a statement in explanation or mitigation.
Failure to provide a statement, or to provide satisfactory evidence to corroborate his/her
explanation, will result in the applicant... | UCL Academic Manual |
F
∈ [0, 1]
(4)
Ar=8
represents the columns of UAr=8 corresponding to the top-i singular vectors.
where U i
φ(·) has a range of [0, 1], where 1 represents a complete overlap of subspaces and 0 a complete
separation. See Figure 3 for how φ changes as we vary i and j. We only look at the 48th layer
(out of 96) due to... | LORA |
Other activities that have been associated with political disinformation
campaigns in the past would potentially run afoul of a host of laws, including
statutes against cyberbullying and the tort of intentional infliction of emotional
distress.56 Insofar as a disinformation campaign made an effort to acquire and
leak in... | Social_Media_and_Democracy |
underlying image representation is made stronger. On au-
dio classification and retrieval benchmarks, IMAGEBIND’s
emergent zero-shot classification matches or outperforms
specialist models trained with direct audio-text supervision
on benchmarks like ESC, Clotho, AudioCaps. IMAGEBIND
representations also outperform spe... | IMAGEBIND- One Embedding Space To Bind Them A |
32
Healthcare
Magic
what causes itchy
rash with discharge
behind the ears?
suggest treatment for
itchy rashes behind
ears. | BiomedGPT |
Coherence This is measured by the similarity between the embedding of generated speech and
that of the audio context, where different embedding models would reflect coherence of different
attributes. VALL-E proposed to use WavLM-TDCNN speaker embedding model [Chen et al., 2022],
which maps an audio clip to a fixed dime... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
GPT models are often trained in two stages. First, they are trained, using a large dataset of text
from the Internet, to predict the next word. The models are then fine-tuned with additional data,
using an algorithm called reinforcement learning from human feedback (RLHF), to produce outputs
that are preferred by human ... | gpt-4-system-card |
• develop metrics to estimate the level of privacy of information exchanged, volunteered by the
user and pried by the system, in each communication flow;
• develop metrics to estimate the level of privacy when cross referencing information exchanged
within more than a single flow.
Where a flow can be seen as ... | informatics-phd-projects-2022-23 |
We can also notice that models with a small number of layers have a hard time staying in context, even if they
do manage to produce syntactically correct English. This suggests that the model lacks the ability to capture the
long-term dependencies and the structure of the story. On the other hand, models with more laye... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
6.1. Impact Assessment
We develop model impact assessments to identify, assess, and document key downstream societal
benefits and harms associated with the development of advanced Gemini models. These are informed
by prior academic literature on language model risks (Weidinger et al., 2021), findings from similar
prior... | gemini_1_report |
Noam Wies, Yoav Levine, Daniel Jannai, and Amnon Shashua. Which transformer architecture fits
In Proceedings of the 38th International
my data? A vocabulary bottleneck in self-attention.
Conference on Machine Learning, pp. 11170–11181. PMLR, July 2021. URL https://proc
eedings.mlr.press/v139/wies21a.html.
Rachel Wilka,... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
5.3.3 Bottlenecks on usefulness
The benefits of deployment listed in the box above—profit, power, prestige, solving social problems,
etc—all require the APS system, once deployed, to be useful in various ways. If such a system
136See Askell et al (2019, p. 9)’s discussion of pharmaceuticals; and see Hunt (2020) on the ... | Is Power-Seeking AI an Existential Risk? |
Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler
Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon,
Nouha Dziri, Shrimai Prabhumoye, Yiming Yang,
Shashank Gupta, Bodhisattwa Prasad Majumder,
Katherine Hermann, Sean Welleck, Amir Yazdan-
bakhsh, and Peter Clark. 2023. Self-refine: Iterative
refinement with self-feedback.
N... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
the categories. Intermediate algebra and precalculus can only be solved with a low accuracy rate
of around 20%. ChatGPT is not good at answering questions on topics including derivatives and | ASurveyonEvaluationofLargeLanguageModels |
Animal-onlyRandom(diversity↓)(diversity↑)Location-onlyRandom(diversity↓)(diversity↑)Creative data diversity (4T1)Condtion diversity (4T1)30.525302032.531.5 that one-round self-refinement effectively utilizes the existing diversity in S and creative data. Consequently, multiple rounds
of self-refinement do not yield a s... | Let’sThinkOutsidetheBox |
Figure 2: The prompt generator architecture. A
T5-base encoder (Raffel et al., 2019) receives train-
able prompt tokens p(cid:48) and the input x, and a cross
attention network implemented following Jaegle
et al. (2021) translates its variable length output se-
quence into a fixed length input dependent prompt,
p(x). Bl... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
space [41] containing all possible input embeddings to the
text encoder. Alternatively, to better capture the target con-
cept, Ruiz et al. [32] proposed the personalization-by-fine-
tuning approach, where one directly fine-tunes the genera-
tive model to represent the user-specified concept. While
this results in bett... | A Neural Space-Time Representation for Text-to-Image Personalization |
in their sample accounted for roughly 80 percent of the potential fake news
exposures that they identified. Regardless of the data or approach, it appears | Social_Media_and_Democracy |
workflow, thereby simplifying these intricate processes. By
curating a corpus infused with domain knowledge and lever-
aging the methodologies offered, one can adeptly fine-tune an
embedding model to align closely with the specific require-
ments of the target domain. | RAG forLargeLanguageModels-ASurvey |
Social simulation can be categorized into macro-level simulation and micro-level simulation [518].
In the macro-level simulation, also known as system-based simulation, researchers model the overall
state of the system of the simulated society [546; 547]. While micro-level simulation, also known as
agent-based simulati... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Flan-PaLM achieves state-of-the-art performance on several benchmarks, such as 75.2% on five-shot MMLU.
Flan-PaLM also has improved usability—for example, it can perform zero-shot reasoning without prompt
engineering or few-shot exemplars. Additionally, we show that instruction finetuning is compatible with
a range of mo... | Scaling Instruction-Finetuned Language Models |
information sharing, management, collection, processing and learning in AI personal assistants. Based
on this, you will design novel methods to personalise privacy in AI assistants based on the social norms
but also on the users' contextual, group, and individual preferences with an optimal accuracy-
intervention tra... | informatics-phd-projects-2022-23 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.